The use of U-Net and Radon transforms for multiple attenuation

Paloma Lira Fontes, Daniel O. Trad, Ivan Sanchez

Radon transform (RT) allows the mapping of multiple and primary reflection events separately in the transformed domain. Hyperbolic Radon transform (HRT) is an example of RT that maps nearly hyperbolic events in the data space to points in the HR space. A methodology of multiple prediction is proposed based on U-Net, a convolutional neural network (CNN) architecture. This network is often applied to image segmentation for classification problems, but the proposed workflow uses the U-Net to predict multiples using HR panels. In this report, we performed predictions using one or two input channels, sparse and nonsparse HR panels, with nonsparse HR panels of multiples as the label. These numerical experiments show that a U-Net can be used to separate the primaries and multiples in the Radon space and therefore predict multiples. This result was achieved using simple geologic models, but further work is required with more complex geologic models. A challenging aspect of this problem is that the transform generates artifacts that are very dependent on the geometry of the input (truncation and sampling artifacts). Because these are very difficult to predict at inference time, they cause a decrease in generalization power.